Building in the open: a company run by AI employees

Update, 28 September 2026: early access is no longer free. Every new workspace starts with ₹25 of free credits and no card; after that you pay as you go for what your AI team uses, in rupees plus GST, with a tax invoice for every top-up. There are no monthly plans. See pricing.

Update, later on 17 September 2026: our AI employees now handle support, sales and product messages and write a daily report, and the public company page is live at dhanurai.com/company. A new direction for Dhanur AI has the details.

We are asking Indian businesses to trust AI employees with their customers, their leads and their money. That is a big ask, and we do not think a landing page earns it. So we are going to show you instead.

Dhanur AI is being built to run in the open. The company's own work is being handed to AI employees that run on Dhanur AI itself, supervised by our founder. We will publish how that goes: what they do, what it costs, and where they get it wrong.

This post explains the plan, the guardrails, and the line we will not cross.

Why run a company this way

There are three reasons.

It is the most honest demo we can give. Any product looks good in a screenshot. Our own support inbox on a bad day, or our own finance checks at the end of the month, show much more. If our AI employees cannot handle that work, you should know before you rely on them for yours.

It keeps us honest about cost. Every AI task has a price, and it is easy to hide it inside a monthly fee. When our own AI employees' spending is on show, we have every reason to keep it low and to show you the same numbers inside the product.

We find the problems first. Each time one of our AI employees hands a task to a person, gives a weak answer or reaches a spending cap, we learn something. It is far better for us to learn it on our own work than on yours.

The roles we are handing over

Here are the roles. Most of them now have real, recurring work; each one says what it does today and what is still to come.

• Support answers customer questions and handles routine account requests. It hands complaints, anything that sounds legal, and refunds above a set limit to the founder. It also sweeps our own inbox every quarter of an hour: a message nobody has drafted an answer for, or a reply that could not be delivered, goes to the founder rather than sitting there.
• Platform operations watches background jobs that failed, a queue that stopped draining, runs that ended in an error, deliveries that failed, the day's AI spending against its cap, and settings this deployment needs and does not have. It raises one item for the founder when something is wrong and says nothing when nothing is. Switching to a backup model automatically is still to come.
• Finance reads yesterday's provider costs each morning, split by AI employee and by workspace, against what we would charge for them, and flags a workspace whose cost jumped or a day that came close to the spending cap. GST reconciliation is still to come.
• Product turns what customers write into roadmap items, grouping new evidence under an idea we already have rather than creating it twice, and once a week reads every conversation together to find what people kept asking about — with how many people asked and from how many businesses, never who they were.
• Trust and safety looks for abuse, such as bulk spam, scraping or impersonation. It also handles data requests under India's Digital Personal Data Protection Act. Suspending an account always needs a human decision. This role is still to come.
• Chief of staff sorts the founder's inbox and writes the daily report: what we shipped that day, taken from the changelog, and what customers did, taken from our own dashboard. A quiet day is published as a quiet day.

What we will publish

We are planning a public company dashboard. It is not live yet, and we will announce it on this blog when it is. It will show:

• Activity. A running feed of what our AI employees did, stripped of anything that identifies a person or a customer.
• Money. Revenue and what we spend on AI providers, either as exact figures or in bands.
• Platform numbers. Tasks run, uptime, and the average cost of a task.
• The team. Each AI employee, its role and what it is allowed to do.

On this blog we will write up what we learn, including the failures: an answer that went wrong and why, a cap that fired, a job we took back from an AI employee because a person does it better.

The guardrails

The same controls we are building for customers apply to our own AI employees.

Approvals. Anything that sends a message to a customer, spends money or touches something marked as sensitive waits for a person to say yes. A role earns more freedom by getting routine work right. Money, legal matters and account decisions stay behind an approval for good.

Spending caps. Each AI employee has a budget per task and per day, and every call to an AI model is recorded in rupees against the employee and the task. At its cap an employee stops; it does not quietly keep going.

A kill switch. One control pauses every AI employee at once. It is built to flip on by itself if the company's daily spending limit is breached, and to alert the founder when that happens.

Human escalation. Each role has written rules for when to hand over to a person: complaints, legal language, refunds above the limit, suspensions, and anything the employee is unsure about. The founder reads escalations first and makes the call.

Tool permissions. An AI employee can only use the tools on its own list. If a tool that moves money is not on the support employee's list, it cannot move money, however a conversation goes.

Disclosure. When one of our AI employees talks to you, it says it is an AI. It will never pretend to be a person.

What we will not publish

Running in the open does not mean exposing the people who use Dhanur AI. We will never publish:

• customer names, phone numbers, email addresses or any other personal details;
• the content of anyone's messages, documents or tasks;
• which businesses use Dhanur AI, unless a business asks us to name it;
• details that would help someone attack the platform or its users.

Public pages will only be able to show data that has passed through a sanitising step built for this purpose. It lets through totals and anonymised activity, and nothing else. Before the dashboard goes live, it will be tested by planting fake personal details in the data and failing the check if any of them reach a public page. If a figure could identify someone, we leave it out or show it as a band.

Your own data is covered by our privacy policy. We do not sell it, and we do not use it to train AI models.

Follow along

Product updates go in the changelog, and these notes go on this blog, which also has an RSS feed. If you would like to see an AI employee at work on your own business, create a workspace and be first in line for the agent builder.

And if you think we have drawn a line in the wrong place, about what we publish or what we keep private, write to outreach@prodigalai.com. We would rather hear it now.